blob: 6a3745c0582ae2388f61ef099f4ecc027427e5da [file] [log] [blame]
// Copyright 2023 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// Bisect finds changes responsible for causing a failure.
// A typical use is to identify the source locations in a program
// that are miscompiled by a given compiler optimization.
// Usage:
// bisect [flags] [var=value...] command [arguments...]
// Bisect operates on a target command line – the target – that can be
// run with various changes individually enabled or disabled. With none
// of the changes enabled, the target is known to succeed (exit with exit
// code zero). With all the changes enabled, the target is known to fail
// (exit any other way). Bisect repeats the target with different sets of
// changes enabled, using binary search to find (non-overlapping) minimal
// change sets that provoke the failure.
// The target must cooperate with bisect by accepting a change pattern
// and then enabling and reporting the changes that match that pattern.
// The change pattern is passed to the target by substituting it anywhere
// the string PATTERN appears in the environment values or the command
// arguments. For each change that matches the pattern, the target must
// enable that change and also print one or more “match lines”
// (to standard output or standard error) describing the change.
// The [] package provides functions to help
// targets implement this protocol. We plan to publish that package
// in a non-internal location after finalizing its API.
// Bisect starts by running the target with no changes enabled and then
// with all changes enabled. It expects the former to succeed and the latter to fail,
// and then it will search for the minimal set of changes that must be enabled
// to provoke the failure. If the situation is reversed – the target fails with no
// changes enabled and succeeds with all changes enabled – then bisect
// automatically runs in reverse as well, searching for the minimal set of changes
// that must be disabled to provoke the failure.
// Bisect prints tracing logs to standard error and the minimal change sets
// to standard output.
// # Command Line Flags
// Bisect supports the following command-line flags:
// -max=M
// Stop after finding M minimal change sets. The default is no maximum, meaning to run until
// all changes that provoke a failure have been identified.
// -maxset=S
// Disallow change sets larger than S elements. The default is no maximum.
// -timeout=D
// If the target runs for longer than duration D, stop the target and interpret that as a failure.
// The default is no timeout.
// -count=N
// Run each trial N times (default 2), checking for consistency.
// -v
// Print verbose output, showing each run and its match lines.
// In addition to these general flags,
// bisect supports a few “shortcut” flags that make it more convenient
// to use with specific targets.
// -compile=<rewrite>
// This flag is equivalent to adding an environment variable
// “GOCOMPILEDEBUG=<rewrite>hash=PATTERN”,
// which, as discussed in more detail in the example below,
// allows bisect to identify the specific source locations where the
// compiler rewrite causes the target to fail.
// -godebug=<name>=<value>
// This flag is equivalent to adding an environment variable
// “GODEBUG=<name>=<value>#PATTERN”,
// which allows bisect to identify the specific call stacks where
// the changed [GODEBUG setting] value causes the target to fail.
// # Example
// The Go compiler provides support for enabling or disabling certain rewrites
// and optimizations to allow bisect to identify specific source locations where
// the rewrite causes the program to fail. For example, to bisect a failure caused
// by the new loop variable semantics:
// bisect go test -gcflags=all=-d=loopvarhash=PATTERN
// The -gcflags=all= instructs the go command to pass the -d=... to the Go compiler
// when compiling all packages. Bisect varies PATTERN to determine the minimal set of changes
// needed to reproduce the failure.
// The go command also checks the GOCOMPILEDEBUG environment variable for flags
// to pass to the compiler, so the above command is equivalent to:
// bisect GOCOMPILEDEBUG=loopvarhash=PATTERN go test
// Finally, as mentioned earlier, the -compile flag allows shortening this command further:
// bisect -compile=loopvar go test
// # Defeating Build Caches
// Build systems cache build results, to avoid repeating the same compilations
// over and over. When using a cached build result, the go command (correctly)
// reprints the cached standard output and standard error associated with that
// command invocation. (This makes commands like 'go build -gcflags=-S' for
// printing an assembly listing work reliably.)
// Unfortunately, most build systems, including Bazel, are not as careful
// as the go command about reprinting compiler output. If the compiler is
// what prints match lines, a build system that suppresses compiler
// output when using cached compiler results will confuse bisect.
// To defeat such build caches, bisect replaces the literal text “RANDOM”
// in environment values and command arguments with a random 64-bit value
// during each invocation. The Go compiler conveniently accepts a
// -d=ignore=... debug flag that ignores its argument, so to run the
// previous example using Bazel, the invocation is:
// bazel test --define=gc_goopts=-d=loopvarhash=PATTERN,unused=RANDOM //path/to:test
// [GODEBUG setting]:
package main
import (
// Preserve import of bisect, to allow [bisect.Match] in the doc comment.
var _ bisect.Matcher
func usage() {
fmt.Fprintf(os.Stderr, "usage: bisect [flags] [var=value...] command [arguments...]\n")
func main() {
log.SetPrefix("bisect: ")
var b Bisect
b.Stdout = os.Stdout
b.Stderr = os.Stderr
flag.IntVar(&b.Max, "max", 0, "stop after finding `m` failing change sets")
flag.IntVar(&b.MaxSet, "maxset", 0, "do not search for change sets larger than `s` elements")
flag.DurationVar(&b.Timeout, "timeout", 0, "stop target and consider failed after duration `d`")
flag.IntVar(&b.Count, "count", 2, "run target `n` times for each trial")
flag.BoolVar(&b.Verbose, "v", false, "enable verbose output")
env := ""
envFlag := ""
flag.Func("compile", "bisect source locations affected by Go compiler `rewrite` (fma, loopvar, ...)", func(value string) error {
if envFlag != "" {
return fmt.Errorf("cannot use -%s and -compile", envFlag)
envFlag = "compile"
env = "GOCOMPILEDEBUG=" + value + "hash=PATTERN"
return nil
flag.Func("godebug", "bisect call stacks affected by GODEBUG setting `name=value`", func(value string) error {
if envFlag != "" {
return fmt.Errorf("cannot use -%s and -godebug", envFlag)
envFlag = "godebug"
env = "GODEBUG=" + value + "#PATTERN"
return nil
flag.Usage = usage
args := flag.Args()
// Split command line into env settings, command name, args.
i := 0
for i < len(args) && strings.Contains(args[i], "=") {
if i == len(args) {
b.Env, b.Cmd, b.Args = args[:i], args[i], args[i+1:]
if env != "" {
b.Env = append([]string{env}, b.Env...)
// Check that PATTERN is available for us to vary.
found := false
for _, e := range b.Env {
if _, v, _ := strings.Cut(e, "="); strings.Contains(v, "PATTERN") {
found = true
for _, a := range b.Args {
if strings.Contains(a, "PATTERN") {
found = true
if !found {
log.Fatalf("no PATTERN in target environment or args")
if !b.Search() {
// A Bisect holds the state for a bisect invocation.
type Bisect struct {
// Env is the additional environment variables for the command.
// PATTERN and RANDOM are substituted in the values, but not the names.
Env []string
// Cmd is the command (program name) to run.
// PATTERN and RANDOM are not substituted.
Cmd string
// Args is the command arguments.
// PATTERN and RANDOM are substituted anywhere they appear.
Args []string
// Command-line flags controlling bisect behavior.
Max int // maximum number of sets to report (0 = unlimited)
MaxSet int // maximum number of elements in a set (0 = unlimited)
Timeout time.Duration // kill target and assume failed after this duration (0 = unlimited)
Count int // run target this many times for each trial and give up if flaky (min 1 assumed; default 2 on command line set in main)
Verbose bool // print long output about each trial (only useful for debugging bisect itself)
// State for running bisect, replaced during testing.
// Failing change sets are printed to Stdout; all other output goes to Stderr.
Stdout io.Writer // where to write standard output (usually os.Stdout)
Stderr io.Writer // where to write standard error (usually os.Stderr)
TestRun func(env []string, cmd string, args []string) (out []byte, err error) // if non-nil, used instead of exec.Command
// State maintained by Search.
// By default, Search looks for a minimal set of changes that cause a failure when enabled.
// If Disable is true, the search is inverted and seeks a minimal set of changes that
// cause a failure when disabled. In this case, the search proceeds as normal except that
// each pattern starts with a !.
Disable bool
// SkipDigits is the number of hex digits to use in skip messages.
// If the set of available changes is the same in each run, as it should be,
// then this doesn't matter: we'll only exclude suffixes that uniquely identify
// a given change. But for some programs, especially bisecting runtime
// behaviors, sometimes enabling one change unlocks questions about other
// changes. Strictly speaking this is a misuse of bisect, but just to make
// bisect more robust, we use the y and n runs to create an estimate of the
// number of bits needed for a unique suffix, and then we round it up to
// a number of hex digits, with one extra digit for good measure, and then
// we always use that many hex digits for skips.
SkipHexDigits int
// Add is a list of suffixes to add to every trial, because they
// contain changes that are necessary for a group we are assembling.
Add []string
// Skip is a list of suffixes that uniquely identify changes to exclude from every trial,
// because they have already been used in failing change sets.
// Suffixes later in the list may only be unique after removing
// the ones earlier in the list.
// Skip applies after Add.
Skip []string
// A Result holds the result of a single target trial.
type Result struct {
Success bool // whether the target succeeded (exited with zero status)
Cmd string // full target command line
Out string // full target output (stdout and stderr combined)
Suffix string // the suffix used for collecting MatchIDs, MatchText, and MatchFull
MatchIDs []uint64 // match IDs enabled during this trial
MatchText []string // match reports for the IDs, with match markers removed
MatchFull []string // full match lines for the IDs, with match markers kept
// &searchFatal is a special panic value to signal that Search failed.
// This lets us unwind the search recursion on a fatal error
// but have Search return normally.
var searchFatal int
// Search runs a bisect search according to the configuration in b.
// It reports whether any failing change sets were found.
func (b *Bisect) Search() bool {
defer func() {
// Recover from panic(&searchFatal), implicitly returning false from Search.
// Re-panic on any other panic.
if e := recover(); e != nil && e != &searchFatal {
// Run with no changes and all changes, to figure out which direction we're searching.
// The goal is to find the minimal set of changes to toggle
// starting with the state where everything works.
// If "no changes" succeeds and "all changes" fails,
// we're looking for a minimal set of changes to enable to provoke the failure
// (broken = runY, b.Negate = false)
// If "no changes" fails and "all changes" succeeds,
// we're looking for a minimal set of changes to disable to provoke the failure
// (broken = runN, b.Negate = true).
b.Logf("checking target with all changes disabled")
runN := b.Run("n")
b.Logf("checking target with all changes enabled")
runY := b.Run("y")
var broken *Result
switch {
case runN.Success && !runY.Success:
b.Logf("target succeeds with no changes, fails with all changes")
b.Logf("searching for minimal set of enabled changes causing failure")
broken = runY
b.Disable = false
case !runN.Success && runY.Success:
b.Logf("target fails with no changes, succeeds with all changes")
b.Logf("searching for minimal set of disabled changes causing failure")
broken = runN
b.Disable = true
case runN.Success && runY.Success:
b.Fatalf("target succeeds with no changes and all changes")
case !runN.Success && !runY.Success:
b.Fatalf("target fails with no changes and all changes")
// Compute minimum number of bits needed to distinguish
// all the changes we saw during N and all the changes we saw during Y.
b.SkipHexDigits = skipHexDigits(runN.MatchIDs, runY.MatchIDs)
// Loop finding and printing change sets, until none remain.
found := 0
for {
// Find set.
bad :=
if bad == nil {
if found == 0 {
b.Fatalf("cannot find any failing change sets of size ≤ %d", b.MaxSet)
// Confirm that set really does fail, to avoid false accusations.
// Also asking for user-visible output; earlier runs did not.
b.Logf("confirming failing change set")
b.Add = append(b.Add[:0], bad...)
broken = b.Run("v")
if broken.Success {
b.Logf("confirmation run succeeded unexpectedly")
b.Add = b.Add[:0]
// Print confirmed change set.
b.Logf("FOUND failing change set")
desc := "(enabling changes causes failure)"
if b.Disable {
desc = "(disabling changes causes failure)"
fmt.Fprintf(b.Stdout, "--- change set #%d %s\n%s\n---\n", found, desc, strings.Join(broken.MatchText, "\n"))
// Stop if we've found enough change sets.
if b.Max > 0 && found >= b.Max {
// If running bisect target | tee bad.txt, prints to stdout and stderr
// both appear on the terminal, but the ones to stdout go through tee
// and can take a little bit of extra time. Sleep 1 millisecond to give
// tee time to catch up, so that its stdout print does not get interlaced
// with the stderr print from the next b.Log message.
time.Sleep(1 * time.Millisecond)
// Disable the now-known-bad changes and see if any failures remain.
b.Logf("checking for more failures")
b.Skip = append(bad, b.Skip...)
broken = b.Run("")
if broken.Success {
what := "enabled"
if b.Disable {
what = "disabled"
b.Logf("target succeeds with all remaining changes %s", what)
b.Logf("target still fails; searching for more bad changes")
return true
// Fatalf prints a message to standard error and then panics,
// causing Search to return false.
func (b *Bisect) Fatalf(format string, args ...any) {
s := fmt.Sprintf("bisect: fatal error: "+format, args...)
if !strings.HasSuffix(s, "\n") {
s += "\n"
// Logf prints a message to standard error.
func (b *Bisect) Logf(format string, args ...any) {
s := fmt.Sprintf("bisect: "+format, args...)
if !strings.HasSuffix(s, "\n") {
s += "\n"
func skipHexDigits(idY, idN []uint64) int {
var all []uint64
seen := make(map[uint64]bool)
for _, x := range idY {
seen[x] = true
all = append(all, x)
for _, x := range idN {
if !seen[x] {
seen[x] = true
all = append(all, x)
sort.Slice(all, func(i, j int) bool { return bits.Reverse64(all[i]) < bits.Reverse64(all[j]) })
digits := sort.Search(64/4, func(digits int) bool {
mask := uint64(1)<<(4*digits) - 1
for i := 0; i+1 < len(all); i++ {
if all[i]&mask == all[i+1]&mask {
return false
return true
if digits < 64/4 {
return digits
// search searches for a single locally minimal change set.
// Invariant: r describes the result of r.Suffix + b.Add, which failed.
// (There's an implicit -b.Skip everywhere here. b.Skip does not change.)
// We want to extend r.Suffix to preserve the failure, working toward
// a suffix that identifies a single change.
func (b *Bisect) search(r *Result) []string {
// The caller should be passing in a failure result that we diagnose.
if r.Success {
b.Fatalf("internal error: unexpected success") // mistake by caller
// If the failure reported no changes, the target is misbehaving.
if len(r.MatchIDs) == 0 {
b.Fatalf("failure with no reported changes:\n\n$ %s\n%s\n", r.Cmd, r.Out)
// If there's one matching change, that's the one we're looking for.
if len(r.MatchIDs) == 1 {
return []string{fmt.Sprintf("x%0*x", b.SkipHexDigits, r.MatchIDs[0]&(1<<(4*b.SkipHexDigits)-1))}
// If the suffix we were tracking in the trial is already 64 bits,
// either the target is bad or bisect itself is buggy.
if len(r.Suffix) >= 64 {
b.Fatalf("failed to isolate a single change with very long suffix")
// We want to split the current matchIDs by left-extending the suffix with 0 and 1.
// If all the matches have the same next bit, that won't cause a split, which doesn't
// break the algorithm but does waste time. Avoid wasting time by left-extending
// the suffix to the longest suffix shared by all the current match IDs
// before adding 0 or 1.
suffix := commonSuffix(r.MatchIDs)
if !strings.HasSuffix(suffix, r.Suffix) {
b.Fatalf("internal error: invalid common suffix") // bug in commonSuffix
// Run 0suffix and 1suffix. If one fails, chase down the failure in that half.
r0 := b.Run("0" + suffix)
if !r0.Success {
r1 := b.Run("1" + suffix)
if !r1.Success {
// suffix failed, but 0suffix and 1suffix succeeded.
// Assuming the target isn't flaky, this means we need
// at least one change from 0suffix AND at least one from 1suffix.
// We are already tracking N = len(b.Add) other changes and are
// allowed to build sets of size at least 1+N (or we shouldn't be here at all).
// If we aren't allowed to build sets of size 2+N, give up this branch.
if b.MaxSet > 0 && 2+len(b.Add) > b.MaxSet {
return nil
// Adding all matches for 1suffix, recurse to narrow down 0suffix.
old := len(b.Add)
b.Add = append(b.Add, "1"+suffix)
r0 = b.Run("0" + suffix)
if r0.Success {
// 0suffix + b.Add + 1suffix = suffix + b.Add is what r describes, and it failed.
b.Fatalf("target fails inconsistently")
bad0 :=
if bad0 == nil {
// Search failed due to MaxSet limit.
return nil
b.Add = b.Add[:old]
// Adding the specific match we found in 0suffix, recurse to narrow down 1suffix.
b.Add = append(b.Add[:old], bad0...)
r1 = b.Run("1" + suffix)
if r1.Success {
// 1suffix + b.Add + bad0 = bad0 + b.Add + 1suffix is what reported as a failure.
b.Fatalf("target fails inconsistently")
bad1 :=
if bad1 == nil {
// Search failed due to MaxSet limit.
return nil
b.Add = b.Add[:old]
// bad0 and bad1 together provoke the failure.
return append(bad0, bad1...)
// Run runs a set of trials selecting changes with the given suffix,
// plus the ones in b.Add and not the ones in b.Skip.
// The returned result's MatchIDs, MatchText, and MatchFull
// only list the changes that match suffix.
// When b.Count > 1, Run runs b.Count trials and requires
// that they all succeed or they all fail. If not, it calls b.Fatalf.
func (b *Bisect) Run(suffix string) *Result {
out :=
for i := 1; i < b.Count; i++ {
r :=
if r.Success != out.Success {
b.Fatalf("target fails inconsistently")
return out
// run runs a single trial for Run.
func (b *Bisect) run(suffix string) *Result {
random := fmt.Sprint(rand.Uint64())
// Accept suffix == "v" to mean we need user-visible output.
visible := ""
if suffix == "v" {
visible = "v"
suffix = ""
// Construct change ID pattern.
var pattern string
if suffix == "y" || suffix == "n" {
pattern = suffix
suffix = ""
} else {
var elem []string
if suffix != "" {
elem = append(elem, "+", suffix)
for _, x := range b.Add {
elem = append(elem, "+", x)
for _, x := range b.Skip {
elem = append(elem, "-", x)
pattern = strings.Join(elem, "")
if pattern == "" {
pattern = "y"
if b.Disable {
pattern = "!" + pattern
pattern = visible + pattern
// Construct substituted env and args.
env := make([]string, len(b.Env))
for i, x := range b.Env {
k, v, _ := strings.Cut(x, "=")
env[i] = k + "=" + replace(v, pattern, random)
args := make([]string, len(b.Args))
for i, x := range b.Args {
args[i] = replace(x, pattern, random)
// Construct and log command line.
// There is no newline in the log print.
// The line will be completed when the command finishes.
cmdText := strings.Join(append(append(env, b.Cmd), args...), " ")
fmt.Fprintf(b.Stderr, "bisect: run: %s...", cmdText)
// Run command with args and env.
var out []byte
var err error
if b.TestRun != nil {
out, err = b.TestRun(env, b.Cmd, args)
} else {
ctx := context.Background()
if b.Timeout != 0 {
var cancel context.CancelFunc
ctx, cancel = context.WithTimeout(ctx, b.Timeout)
defer cancel()
cmd := exec.CommandContext(ctx, b.Cmd, args...)
cmd.Env = append(os.Environ(), env...)
// Set up cmd.Cancel, cmd.WaitDelay on Go 1.20 and later
// TODO(rsc): Inline go120.go's cmdInterrupt once we stop supporting Go 1.19.
out, err = cmd.CombinedOutput()
// Parse output to construct result.
r := &Result{
Suffix: suffix,
Success: err == nil,
Cmd: cmdText,
Out: string(out),
// Calculate bits, mask to identify suffix matches.
var bits, mask uint64
if suffix != "" && suffix != "y" && suffix != "n" && suffix != "v" {
var err error
bits, err = strconv.ParseUint(suffix, 2, 64)
if err != nil {
b.Fatalf("internal error: bad suffix")
mask = uint64(1<<len(suffix)) - 1
// Process output, collecting match reports for suffix.
have := make(map[uint64]bool)
all := r.Out
for all != "" {
var line string
line, all, _ = strings.Cut(all, "\n")
short, id, ok := bisect.CutMarker(line)
if !ok || (id&mask) != bits {
if !have[id] {
have[id] = true
r.MatchIDs = append(r.MatchIDs, id)
r.MatchText = append(r.MatchText, short)
r.MatchFull = append(r.MatchFull, line)
// Finish log print from above, describing the command's completion.
if err == nil {
fmt.Fprintf(b.Stderr, " ok (%d matches)\n", len(r.MatchIDs))
} else {
fmt.Fprintf(b.Stderr, " FAIL (%d matches)\n", len(r.MatchIDs))
if err != nil && len(r.MatchIDs) == 0 {
b.Fatalf("target failed without printing any matches\n%s", r.Out)
// In verbose mode, print extra debugging: all the lines with match markers.
if b.Verbose {
b.Logf("matches:\n%s", strings.Join(r.MatchFull, "\n\t"))
return r
// replace returns x with literal text PATTERN and RANDOM replaced by pattern and random.
func replace(x, pattern, random string) string {
x = strings.ReplaceAll(x, "PATTERN", pattern)
x = strings.ReplaceAll(x, "RANDOM", random)
return x
// commonSuffix returns the longest common binary suffix shared by all uint64s in list.
// If list is empty, commonSuffix returns an empty string.
func commonSuffix(list []uint64) string {
if len(list) == 0 {
return ""
b := list[0]
n := 64
for _, x := range list {
for x&((1<<n)-1) != b {
b &= (1 << n) - 1
s := make([]byte, n)
for i := n - 1; i >= 0; i-- {
s[i] = '0' + byte(b&1)
b >>= 1
return string(s[:])